On Language Models' Sensitivity to Suspicious Coincidences